So you’re considering AI Safety
This guide is more specific than my post on taste and supersedes my post on transitioning to AI policy. It includes exercises and an opinionated basics reading list, but I recommend ~all hyperlinks here. Nice! Welcome aboard. We’ll need all the help that we can get. There was a period of time I remember so long ago where friends and I used to ask what would it take to get a skeptic to believe? Would it take a real AI system (not specially prompted, or otherwise safeguard-removed), breaking out of a real evaluation environment and hacking into a real commercial company (and, for good measure, collaborating evilly with its fellow AI agents, peer-pressuring the handful of agents who wanted to whistleblow, self-sacrificing for each other, discovering novel zero-days)? Well! May I present to you: The Rise and Fall of Agent Civilizations. I think this’ll give you a little bit of a sense of where the world is today. Here are things I’d do, in roughly this order, and with some words on why: Fi